Pashto & Afghan-Language Speech Recognition Benchmarking

A single word-error-rate is a model's best face, not its real one.

A speech model reports one headline number, and the number is an average. It tells you the model works for the speakers in the benchmark — typically standard dialect, clean audio, prestige speech — and it says nothing about the rest. The same model that scores well on standard Dari mistranscribes Hazaragi, stumbles on a regional accent, and breaks on real-world audio, and the average reports none of it. In low-resource languages the gap is not marginal: word-error-rates stay high precisely because of dialectal variation, and in the worst cases climb past the point where the output is usable at all.

Whose voice the model fails on is the question the average refuses to answer — and the field knows it. Evaluation has moved beyond a single WER toward robustness across speakers, environments, and dialects, because one number was never enough. But measuring that requires what almost no one has: native-validated reference transcriptions across every dialect band, and listeners who can judge synthesis quality in the language as it is actually spoken.

The same is true in reverse. A synthesized voice that speaks a prestige dialect to a population that does not is a voice the population does not recognize — and voice synthesis, now mainstream, raises its own questions of consent that careless deployment ignores.

Ariana Nexus measures what the average hides: independent ASR and TTS evaluation across all 24 Afghan languages and their dialects, word-error-rate broken out by dialect band, native-validated reference data, and the Sovereign Speech Index — with voice data sourced by consent and synthesis handled responsibly.